WALT: New Watermarking Method for Avatar Videos Achieves 92.4% Robustness to Zoom Attacks
A new benchmark called RAW (Robust Avatar Watermarking) has been launched by researchers, featuring 50 synthetic avatar videos sourced from 5 different commercial providers, along with 6 attacks that mimic real-world avatar scenarios. An assessment of 7 current techniques showed that attacks tailored to avatars, like background removal, severely impact watermark retrieval. In response, the team has developed WALT (Watermarking Avatars with Learned Textures), which integrates watermarks within UV texture space through 3D face reconstruction. WALT demonstrates exceptional resilience against zoom attacks with a success rate of 92.4% and performs well on background removal at 95.6%. This benchmark aims to advance research in watermarking specifically for avatars.
Key facts
- RAW benchmark includes 50 synthetic avatar videos from 5 commercial providers.
- 6 attacks simulate real-world avatar workflows.
- 7 existing methods were evaluated.
- Background removal significantly degrades watermark recovery.
- WALT embeds watermarks in UV texture space via 3D face reconstruction.
- WALT achieves 92.4% robustness to zoom attacks.
- WALT achieves 95.6% robustness to background removal.
- The benchmark is released to facilitate research.
Entities
Institutions
- arXiv